Classifying Brand Preference

Rhys Hewer

1 Executive Summary

REMEMBER!! Combine the data sets at the end!

1.1 Objective

1.2 Method

1.3 Findings

1.4 Recommendations

2 Initial Data Processing

#load libraries
library(readxl)
library(dplyr)
library(ggplot2)
library(plotly)
library(RColorBrewer)
library(corrplot)
library(caret)
library(tidyr)
library(kableExtra)
library(parallel)
library(doParallel)
library(rattle)
library(rpart)
#read in data
setwd("C:/Users/rhysh/Google Drive/Data Science/Ubiqum/Project 2/Task 2")
incomplete <- read.csv("SurveyIncomplete.csv")
origdata <- read_xlsx("Survey_Key_and_Complete_Responses_excel.xlsx", sheet = 2)
data <- origdata

As we are working across 2 spreadsheets it is important to check to see if they are structured alike or whether manipulation will be required to ensure they have the same features.

#Check structure of data to see if alike
names(incomplete) == names(data)
## [1] TRUE TRUE TRUE TRUE TRUE TRUE TRUE

The features match between the spreadsheets so no data manipulation is needed in that respect.

incomplete %>% sample_n(5)
##         salary age elevel car zipcode    credit brand
## 270   22229.85  29      4  12       2 437096.86     0
## 2262  99080.50  80      2   2       3  75020.14     0
## 820  106879.76  73      3   6       8 306411.80     0
## 3669  71122.16  48      3  12       4 332028.44     0
## 4830 138649.26  43      2   5       7  23902.55     0
data %>% sample_n(5) 
## # A tibble: 5 x 7
##    salary   age elevel   car zipcode  credit brand
##     <dbl> <dbl>  <dbl> <dbl>   <dbl>   <dbl> <dbl>
## 1 137576.    50      4    15       3   2000.     1
## 2 114157.    33      0    15       8 328351.     1
## 3  92876.    57      3    16       7 232789.     1
## 4 126217.    30      2     3       7 395842.     1
## 5 125200.    66      0     7       2 277273.     1

A quick look at a sample from both spreadsheets shows that they are very similar in composition.

str(incomplete)
## 'data.frame':    5000 obs. of  7 variables:
##  $ salary : num  110500 140894 119160 20000 93956 ...
##  $ age    : int  54 44 49 56 59 71 32 33 32 58 ...
##  $ elevel : int  3 4 2 0 1 2 1 4 1 2 ...
##  $ car    : int  15 20 1 9 15 7 17 17 19 8 ...
##  $ zipcode: int  4 7 3 1 1 2 1 0 2 4 ...
##  $ credit : num  354724 395015 122025 99630 458680 ...
##  $ brand  : int  0 0 0 0 0 0 0 0 0 0 ...
str(data)
## Classes 'tbl_df', 'tbl' and 'data.frame':    10000 obs. of  7 variables:
##  $ salary : num  119807 106880 78021 63690 50874 ...
##  $ age    : num  45 63 23 51 20 56 24 62 29 41 ...
##  $ elevel : num  0 1 0 3 3 3 4 3 4 1 ...
##  $ car    : num  14 11 15 6 14 14 8 3 17 5 ...
##  $ zipcode: num  4 6 2 5 4 3 5 0 0 4 ...
##  $ credit : num  442038 45007 48795 40889 352951 ...
##  $ brand  : num  0 1 0 1 0 1 1 1 0 1 ...

Looking at the structure shows that a few changes are needed in respect of the data types.

Overall, however, the datasets are sufficiently similar. I will take the strategy of splitting into training, test and final prediction sets. The training and testing sets will come from the complete responses data. Final prediction from the incomplete responses.

The exploratory data analysis will be performed on the complete data but any data transformations taking place on the training/testing sets will also need to be made on the Final prediction data prior to modelling this.

#check for NAs
data %>% is.na() %>% sum()
## [1] 0
incomplete %>% is.na() %>% sum()
## [1] 0

There are no missing values in either data set.

#data types
data$elevel <- data$elevel %>% as.factor()
data$car <- data$car %>% as.factor()
data$zipcode <- data$zipcode %>% as.factor()
data$brand <- data$brand %>% as.factor()

Education level, car owned, zip code and brand preference are converted from numeric to factors.

##check for outliers
numericVars <- Filter(is.numeric, data)
outliers <- numericVars %>% sapply(function(x) boxplot(x, plot=FALSE)$out) %>% str()
## List of 3
##  $ salary: num(0) 
##  $ age   : num(0) 
##  $ credit: num(0)

There are no outliers.

3 Exploratory Data Analysis

3.1 Feature Histograms

3.1.1 Brand Preference Histogram

#EDA

##Key feature is brand preference, begin with exploring this value.

g6 <- ggplot(data, aes(brand, fill = brand)) +
        geom_bar() +
        theme_bw() +
        scale_fill_brewer(palette="Dark2") +
        xlab("Brand Preference") + 
        ylab("Frequency") + 
        ggtitle("Brand Preference Frequencies")
g6

  • We see a clear preference for brand 1 over brand 0.

3.1.2 Histogram of other variables

##review other histograms for skewdness
histData <- origdata %>% select(-brand)
g8 <- ggplot(gather(histData), aes(value)) + 
        geom_histogram(bins = 10, fill = "#D95F02", colour = "white") + 
        theme_bw() +
        facet_wrap(~key, scales = 'free_x') +
        xlab("Value") + 
        ylab("Count") + 
        ggtitle("Histograms of Numeric Variables")
g8

  • Across the remaining features we see no extreme skewing or noteworthy patterns.

3.2 Brand Choice Plotting

3.2.1 Brand Choice v Salary

## Plot Brand choice v other variables
g1 <- ggplot(data, aes(brand, salary, fill = brand)) +
        geom_violin() +
        theme_bw() +
        scale_fill_brewer(palette="Dark2") +
        xlab("Brand Preference") + 
        ylab("Salary ($)") + 
        ggtitle("Brand Preference v Salary")
g1 <- ggplotly(g1)
g1
  • There is a clear pattern: salaries ranging between 45k - 100k seem to have a preference for brand 0. Salaries outside of this range seem to have a preference for brand 1.

3.2.2 Brand Choice v Age

g2 <- ggplot(data, aes(brand, age, fill = brand)) +
        geom_violin() +
        theme_bw() +
        scale_fill_brewer(palette="Dark2") +
        xlab("Brand Preference") + 
        ylab("Age") + 
        ggtitle("Brand Preference v Age")
g2 <- ggplotly(g2)
g2
  • There is no noteworthy pattern between brand preference and age.

3.2.3 Brand Choice v Education Level

g3 <- ggplot(data, aes(brand, elevel, colour = brand)) +
        geom_count() +
        theme_bw() +
        scale_color_brewer(palette="Dark2") +
        xlab("Brand Preference") + 
        ylab("Education Level") + 
        ggtitle("Brand Preference v Education Level")
g3

  • The general preference for brand 1 is shown in this plot but there also seems to be a consistent preference for brand 1 regardless of education level whereas there seems to be more variation within preference for brand 0 based on the education level.

3.2.4 Brand Choice v Car

g4 <- ggplot(data, aes(brand, car, colour = brand)) +
        geom_count() +
        theme_bw() +
        scale_color_brewer(palette="Dark2") +
        xlab("Brand Preference") + 
        ylab("Car") + 
        ggtitle("Brand Preference v Car")
g4

  • The general preference for brand 1 is shown in this plot but there also seems to be a consistent preference for brand 1 regardless of car whereas there seems to be more variation within preference for brand 0 based on the car owned.

3.2.5 Brand Choice v Zip Code

g5 <- ggplot(data, aes(brand, zipcode, colour = brand)) +
        geom_count() +
        theme_bw() +
        scale_color_brewer(palette="Dark2") +
        xlab("Brand Preference") + 
        ylab("Zip Code") + 
        ggtitle("Brand Preference v Zip Code")
g5

  • The general preference for brand 1 is shown in this plot but there also seems to be a consistent preference for brand 1 regardless of zip code whereas there seems to be more variation within preference for brand 0 based on the car owned.

3.2.6 Brand Choice v Credit

g7 <- ggplot(data, aes(brand, credit, fill = brand)) +
        geom_violin() +
        theme_bw() +
        scale_fill_brewer(palette="Dark2") +
        xlab("Brand Preference") + 
        ylab("Credit") + 
        ggtitle("Brand Preference v Credit")
g7 <- ggplotly(g7)
g7
  • There is no noteworthy pattern between brand preference and credit.

3.3 Initial Hypothesis

There is a general preference for brand 1 and salary seems the key feature.

Age and credit seem to have very limited influence on brand preference. Whilst there are patterns of preference within education level, car and zip code, salary appears to have the most striking impact.

Salaries ranging between 45k - 100k seem to have a preference for brand 0. Salaries outside of this range seem to have a preference for brand 1.

3.4 Feature Selection

3.4.1 Decision Tree Variable Importance

We can review the initial hypothesis and provide some information on which to make feature selection decisions using a decision tree.

featureDT <- rpart(brand ~ ., data = data)
featureDT$variable.importance
##        age     salary        car    zipcode     credit     elevel 
## 2036.06398 1421.30552   78.59978   42.70644   21.20404   16.06770
fancyRpartPlot(featureDT)

The decision tree is formed exclusively of age and salary, suggesting these are the key features. This is reinforced by the variable importance information from within the model.

3.4.2 Supplementary Exploratory Data Analysis

g12 <- ggplot(data, aes(salary, age, colour = brand)) +
        geom_point(show.legend = FALSE) +
        facet_grid(brand ~ .) +
        theme_bw() +
        scale_color_brewer(palette="Dark2") +
        xlab("Salary") + 
        ylab("Age") + 
        ggtitle("Salary v Age by Brand Preference")
g12

We see that the interaction of age and salary do play a role in brand preference. There are clearly defined blocks of preference depending on age and salary. For example, between ages 60-80 earning between 25k-80k have a distinct preference for brand 0.

I have updated the initial hypothesis to now consider that age and salary are the key features in determining brand preference.

3.4.3 Colinearity and Variance

##Review correlation matrix
corrMatrix <- origdata %>% cor()
corrMatrix %>% corrplot.mixed()

Reviewing the correlation plot we see very little correlation between the features. This means that colinearity is not an issue that needs to be addressed.

#near zero variance
nzv <- data %>% nearZeroVar(saveMetrics = TRUE)
nzv
##         freqRatio percentUnique zeroVar   nzv
## salary   1.112069         97.53   FALSE FALSE
## age      1.178744          0.61   FALSE FALSE
## elevel   1.035946          0.05   FALSE FALSE
## car      1.026415          0.20   FALSE FALSE
## zipcode  1.020105          0.09   FALSE FALSE
## credit   1.057377         97.51   FALSE FALSE
## brand    1.643405          0.02   FALSE FALSE

There are no features with near zero variance.

3.4.4 Feature Selection Conclusion

The Random Forest algorithm has built-in feature selection so no feature selection required from that perspective.

For KNN I will take both a fully-featured dataset and a concentrated dataset (brand, salary, age) into the modelling phase and compare results.

4 Modelling

4.1 Modelling Preparation

#Creating Testing/Training sets
set.seed(111)
trainIndex <- createDataPartition(iris$Species, p = 0.75, list = FALSE)

training <- data[ trainIndex,]
testing  <- data[-trainIndex,]

concData <- data %>% select(brand, age, salary)
training.conc <- concData[ trainIndex,]
testing.conc <- concData[-trainIndex,]

#set up parallel processing (requires parallel and doParallel libraries)
cluster <- makeCluster(detectCores() - 1) 
registerDoParallel(cluster)

#Cross Validation 10 fold
fitControl<- trainControl(method = "cv", number = 10, savePredictions = TRUE, allowParallel = TRUE)

A test and traing set are created with all features. An additional training set is created with just Brand, Age and Salary. 10-fold cross validation is being used and parallel processing is being leveraged.

4.2 Modelling K-Nearest Neighbour

4.2.1 Training KNN

4.2.1.1 KNN - All Features

# Deploy KNN model 
model.KNN.brand <- train(brand ~ ., data = training, 
                         method = "knn", trControl = fitControl)
model.KNN.brand
## k-Nearest Neighbors 
## 
## 114 samples
##   6 predictor
##   2 classes: '0', '1' 
## 
## No pre-processing
## Resampling: Cross-Validated (10 fold) 
## Summary of sample sizes: 102, 103, 103, 102, 103, 102, ... 
## Resampling results across tuning parameters:
## 
##   k  Accuracy   Kappa    
##   5  0.7386364  0.2455904
##   7  0.7196970  0.1720068
##   9  0.7015152  0.1015749
## 
## Accuracy was used to select the optimal model using the largest value.
## The final value used for the model was k = 5.

4.2.1.2 KNN - Concentrated Features

# Deploy KNN model with concentrated feature set
model.KNN.brand.conc <- train(brand ~ ., data = training.conc, 
                              method = "knn", trControl = fitControl)
model.KNN.brand.conc
## k-Nearest Neighbors 
## 
## 114 samples
##   2 predictor
##   2 classes: '0', '1' 
## 
## No pre-processing
## Resampling: Cross-Validated (10 fold) 
## Summary of sample sizes: 103, 103, 103, 103, 102, 103, ... 
## Resampling results across tuning parameters:
## 
##   k  Accuracy   Kappa    
##   5  0.7346154  0.2830504
##   7  0.7451049  0.3028544
##   9  0.7298368  0.2675434
## 
## Accuracy was used to select the optimal model using the largest value.
## The final value used for the model was k = 7.

4.2.1.3 KNN - All Features, Scaled and Centred

# Deploy KNN model with scaled features
model.KNN.brand.scale <- train(brand ~ ., data = training, 
                         method = "knn", trControl = fitControl, 
                         preProcess = c("center", "scale"))
model.KNN.brand.scale
## k-Nearest Neighbors 
## 
## 114 samples
##   6 predictor
##   2 classes: '0', '1' 
## 
## Pre-processing: centered (34), scaled (34) 
## Resampling: Cross-Validated (10 fold) 
## Summary of sample sizes: 103, 103, 102, 102, 103, 103, ... 
## Resampling results across tuning parameters:
## 
##   k  Accuracy   Kappa      
##   5  0.6674242  -0.03808089
##   7  0.6750000  -0.05051582
##   9  0.6931818  -0.04586466
## 
## Accuracy was used to select the optimal model using the largest value.
## The final value used for the model was k = 9.

4.2.1.4 KNN - Concentrated Features, Scaled and Centred

# Deploy KNN model with scaled features and concentrated feature set
model.KNN.brand.conc.scale <- train(brand ~ ., data = training.conc, 
                              method = "knn", trControl = fitControl, 
                              preProcess = c("center", "scale"))
model.KNN.brand.conc.scale
## k-Nearest Neighbors 
## 
## 114 samples
##   2 predictor
##   2 classes: '0', '1' 
## 
## Pre-processing: centered (2), scaled (2) 
## Resampling: Cross-Validated (10 fold) 
## Summary of sample sizes: 103, 103, 103, 103, 102, 101, ... 
## Resampling results across tuning parameters:
## 
##   k  Accuracy   Kappa    
##   5  0.8679487  0.6380495
##   7  0.8581002  0.6001548
##   9  0.8246503  0.4795492
## 
## Accuracy was used to select the optimal model using the largest value.
## The final value used for the model was k = 5.

4.2.1.5 KNN - best model tuned

#tune best KNN model further
tGridKNN <- expand.grid(k = c(1,2,3,4,5,6))
finalKNN <- train(brand ~ ., data = training.conc,
                  method = "knn", trControl = fitControl,
                  preProcess = c("center", "scale"),
                  tuneGrid = tGridKNN)
finalKNN
## k-Nearest Neighbors 
## 
## 114 samples
##   2 predictor
##   2 classes: '0', '1' 
## 
## Pre-processing: centered (2), scaled (2) 
## Resampling: Cross-Validated (10 fold) 
## Summary of sample sizes: 101, 103, 103, 103, 103, 102, ... 
## Resampling results across tuning parameters:
## 
##   k  Accuracy   Kappa    
##   1  0.9113636  0.7810864
##   2  0.9030303  0.7682660
##   3  0.9303030  0.8325011
##   4  0.8688228  0.6379981
##   5  0.9036713  0.7356803
##   6  0.8597319  0.5987915
## 
## Accuracy was used to select the optimal model using the largest value.
## The final value used for the model was k = 3.

By tuning the model I was able to increase the accuracy on the training sent to 91% and the Kappa to 0.78. This brings concerns of overfitting.

4.2.2 Testing KNN

#Predictions on the test set (model = model name, testing = test set)
predictions.KNN <- predict(finalKNN, testing)
testing$predictions.KNN <- predictions.KNN


#Confusion matrix
confMatrix <- confusionMatrix(testing$brand, testing$predictions.KNN)
confMatrix
## Confusion Matrix and Statistics
## 
##           Reference
## Prediction    0    1
##          0 3056  695
##          1  588 5547
##                                           
##                Accuracy : 0.8702          
##                  95% CI : (0.8634, 0.8768)
##     No Information Rate : 0.6314          
##     P-Value [Acc > NIR] : < 2.2e-16       
##                                           
##                   Kappa : 0.7229          
##  Mcnemar's Test P-Value : 0.003083        
##                                           
##             Sensitivity : 0.8386          
##             Specificity : 0.8887          
##          Pos Pred Value : 0.8147          
##          Neg Pred Value : 0.9042          
##              Prevalence : 0.3686          
##          Detection Rate : 0.3091          
##    Detection Prevalence : 0.3794          
##       Balanced Accuracy : 0.8636          
##                                           
##        'Positive' Class : 0               
## 
fourfoldplot(confMatrix$table, conf.level = 0, margin = 1, main = "Confusion Matrix")

The Accuracy of the model on the test set was 87% with a kappa of 0.72. These seem to be satisfactory outcomes. Reviewing the confusion matrix we see that the model is doing a fairly good job of classifying brand preference with a slight tendency to overpredict brand 0.

4.3 Modelling Random Forest

4.3.1 Training Random Forest

4.3.1.1 Random Forest on full feature set

# Deploy RF model 
tGridRF <- expand.grid(mtry = c(24,30,35))
model.RF.brand.tune <- train(brand ~ ., data = training, 
                        method = "rf", trControl = fitControl,
                        tuneGrid = tGridRF)
model.RF.brand.tune
## Random Forest 
## 
## 114 samples
##   6 predictor
##   2 classes: '0', '1' 
## 
## No pre-processing
## Resampling: Cross-Validated (10 fold) 
## Summary of sample sizes: 102, 103, 102, 102, 103, 103, ... 
## Resampling results across tuning parameters:
## 
##   mtry  Accuracy   Kappa    
##   24    0.7734848  0.3644853
##   30    0.7833333  0.4161136
##   35    0.7833333  0.4161136
## 
## Accuracy was used to select the optimal model using the largest value.
## The final value used for the model was mtry = 30.

4.3.1.2 Random Forest on concentrated feature set

## Deploy RF model on concentrated feature set
model.RF.brand.conc <- train(brand ~ ., data = training.conc, 
                         method = "rf", trControl = fitControl)
## note: only 1 unique complexity parameters in default grid. Truncating the grid to 1 .
model.RF.brand.conc 
## Random Forest 
## 
## 114 samples
##   2 predictor
##   2 classes: '0', '1' 
## 
## No pre-processing
## Resampling: Cross-Validated (10 fold) 
## Summary of sample sizes: 103, 103, 103, 103, 103, 103, ... 
## Resampling results:
## 
##   Accuracy  Kappa    
##   0.89662   0.7457665
## 
## Tuning parameter 'mtry' was held constant at a value of 2
varImp(model.RF.brand.conc)
## rf variable importance
## 
##        Overall
## salary     100
## age          0

4.4 Modelling Conclusions

5 Conclusions